100 companies
I thought I was keeping up with the AI industry. Big launches, new tools, staying current.
Then I spent 2 weeks going deep on recent Y Combinator batches: 100+ AI startups that had raised rounds. MRR, team size, founders, what they actually built. Dumped everything into a spreadsheet.
Turns out I had no idea what was actually happening. The numbers first: up to 88% of 2025 batch startups are AI-native. Half are building agents. 80% are B2B. A quarter of W25 wrote 95% of their code with AI tools. Companies hit $10M revenue with teams of under 10 people.
But the numbers aren't the interesting part. The patterns are:
1/ Vertical agents ate horizontal ones
YC no longer bets on AI assistants for everything. The winners are going impossibly narrow:
LunaBill is a voice bot with one job: calling insurance companies about unpaid clinic invoices. Launched in July, $428K in real revenue by year-end, $336K in signed contracts. Every single pilot customer converted to paying. Every one. I used to dismiss companies like this as "just a wrapper." But that's exactly the point. They're not selling technology. They're solving a specific, painful problem in a tiny industry, and they own it.
Vesence founders lived inside a Swedish law firm for months. Built a co-pilot that catches document errors before they go to clients. First firm-wide deployment: 90% of employees use it weekly. Their primary cost of goods? Literally prompts.
2/ Agent infrastructure is the new picks and shovels
When everyone's building agents, someone has to build the plumbing. Memory layers, observability, MCP gateways, eval systems. In YC's Fall 2025 batch alone: 13 startups just in this category.
Stripe for payments. AWS for servers. The same invisible layer is forming for agents — and whoever becomes it inherits the same lock-in. Once you build on someone's agent infra, switching is brutally expensive.
3/ 10 people. $10M revenue.
Garry Tan says it directly: the next $100B companies will have 10 people. The defining metric of this era is revenue per employee.
VideoGen is 2 founders from a dorm room. $30K seed money from Palantir and AWS internship savings. Hit $1M+ ARR before joining YC. Now serving teams at Google and ByteDance. This used to be an anomaly. Now it's a quarter of the batch.
TL;DR
Stop watching the top layer. Model launches and benchmarks are noise. The real game is one level down, in boring niches where tiny teams build products for very specific jobs-to-be-done. Models are becoming commodity. Everything interesting is happening in domain depth.